06-reference

commoncog working backwards

Tue Apr 14 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·reference ·source: Commoncog ·by Cedric Chin

“Book Summary: Working Backwards” — @CedricChin

Why this is in the vault

Chin’s summary of Bryar & Carr’s Amazon-insider book is the operational companion to his first-principles data-driven piece. The money idea — “Good intentions don’t work. Mechanisms do.” — is load-bearing for RDCO’s consulting posture, and the Controllable Input Metrics / WBR section is the same intellectual lineage MAC sits in. Filed as a reference anchor for mechanism-design arguments and for the PR/FAQ format, which RDCO should steal directly for client engagements.

The core argument (paraphrased)

Amazon’s outsized success across disparate industries is explained not by strategy or talent alone but by org design — specifically, a set of explicit leadership principles and five reinforcing mechanisms. Principles describe the culture; mechanisms enforce it when good intentions fail.

The five mechanisms:

  1. Leadership Principles as cultural technology. 14 explicit principles, drilled into every hire, used as rules-of-engagement in every meeting, and — most importantly — used as the design basis for new mechanisms. Bezos: “you’re discovering it, uncovering it — not creating it.”
  2. Bar Raiser hiring. Structured interviews plus a veto-holding interviewer from outside the hiring team whose incentives are decoupled from the urgency to fill the seat. Prevents the slow drift of standards under hiring pressure.
  3. Single-Threaded Leadership. Each major initiative gets one leader with no other responsibilities, running a separable team with minimal dependencies. Requires “services-first” architecture (APIs between teams, not meetings). “Eliminate communication, not encourage it.”
  4. Six-Pager narratives. Replace PowerPoint with dense, prose memos read silently at meeting start. Tufte-inspired. Forces interconnectedness of thought; Bezos “assumes each sentence he reads is wrong until he can prove otherwise.”
  5. PR/FAQ — Working Backwards. Write the press release and customer FAQ before building. Product evaluation skill is passed by osmosis through repeated review cycles — tacit knowledge, not explicable rubric.
  6. Controllable Input Metrics. Not leading indicators, but specifically the controllable inputs the team can move. Every metric has an owner; finance audits independently; WBR reviews them weekly; metrics are retired when they stop measuring what they claim. Goodhart-resistant by design.

Chin’s framing: you cannot copy these wholesale. The mechanisms only work inside the cultural context that generated them. The transferable layer is the meta-mechanism — principles → mechanisms, and the discipline of designing new ones when behavior drifts.

Mapping against Ray Data Co

1. “Good intentions don’t work. Mechanisms do.” is the RDCO consulting thesis. Every client I’ll meet has good intentions about data quality. MAC is a mechanism — it enforces the standard when the deadline pressure hits and the analyst wants to ship the dashboard anyway. Bar Raiser logic applies directly: the person enforcing MAC should have incentives decoupled from the team shipping the model. In a solo practice, that role is me. In a larger client, it’s a gatekeeping function we design into the org. See ../04-tooling/rdco-state-ownership-architecture.

2. PR/FAQ is the format RDCO should steal for engagement scoping. Every consulting engagement should open with a “Working Backwards” memo: here is the press release describing the client’s state 6 months after engagement end. Here is the FAQ. If we can’t write it concretely, we don’t understand the engagement. This is also a better deliverable to the client than a SOW. Add to the playbook. Relevant to 2026-04-14-levie-agent-deployer-role-jd Posture 2 (Playbook + Coaching).

3. Controllable Input Metrics is the RDCO dashboard design principle. When we hand the client a data-quality scorecard, the default temptation is to show output metrics (incidents per month, % passing tests). The Amazon insight is that you track the controllable inputs that produce those — e.g., ”% of models with MAC coverage,” ”% of severity-Stop tests with named owners,” “days since last operational-definition review.” This is the WBR pattern applied to MAC. Direct input into the consulting deliverable template.

4. Single-Threaded Leadership is the honest frame for the phData vs MG decision. The phData role, if it’s a narrow agent-deployer mandate with a separable team, is structurally closer to single-threaded leadership than the MG role (which sounds like a matrixed, dependency-heavy position). That’s an input to the career-moats evaluation. Not decisive — comp delta is real — but it’s a lens the decision doc should name explicitly. See ../01-projects/career-decision-phdata-vs-mg if it exists.

5. “Leadership principles are discovered, not invented.” Relevant to the Sanity Check newsletter voice work. The founder’s voice isn’t something to construct — it’s something to extract from the writing that already works and then enforce mechanically going forward. The voice-match skill is the Bar Raiser for drafts.